Re: Japan built world's fastest computer

[Doyle Saylor <djsaylor@xxxxxxxxxxxxxx>]

Greetings Economists,
On another list we had some discussion of this issue.

Paul starts with,
Jack Dongarra is the Times consultant on High Performance Computing, it
seems. An excellent person, friendly, approachable and extremely busy.
Also he knows a LOT about the mathematical foundations of the algorithms
that simulations depend on. And he is renowned for his nonnumerical
software development ideas.

Supercomputer rankings are by a single parameter, the number of operations
per second, the number of ``flops.'' Even cars are rated using more than
one parameter. Instead of top speed, as an example, people also want to
know the time for acceeration from 0 to 60.

Nevertheless, a general truth about supercomputers: Flop rate is a more
important parameter than automobile top speed.

Now to make a point.

Despite the ovewhelming importance of flops, there are simulation problems
that will not be solved no matter how fast the supercomputer is. A
comparison might be made to transportation problems that will not be
solved no matter what the average automobile top speed is.

Weather forecasting simulation traditionally depends on so-called explicit
algorithms that are ideally suited for vector as well as parallel
machines. This statement oversimplifies.

Generally explicit algorithms reign in supercomputer codes whether it is
forecasting or weapons development. The other option is ``implicit''.
Implicit codes are much more difficult to use. They slow a machine way
down. Poof there goes the high flop rate. The picture compares to driving
a performance automobile in downtown Alameda. The top speed is
unattainable.

Then why use implicit? The answer: Stability and robustness. The objective
is to simulate. If the simulation never completes with explicit and does
with implicit then performance depends on more than just flop rate. Yet it
also depends on flop rate. Implicit is robust but if it takes five years
to run a simulation instead of an hour with explicit (when explicit
works), then

-->  a fast machine is needed not only for explicit <--
-->  algorithms but also for implicit algorithms.   <--


Another point.

Algorithms are the neglected component of simulation. But not the only one
neglected. Another is word length. Some algorithms work better if the word
length increases. In this case, the barrier is not scientific as it is
with algorithms. More algorithm research will improve simulations run on
supercomputers. In the case of the effect of word length, research should
be done on this but even if research shows word length matters, there is
no hardware way to get a longer word length. It can be done with software
but that is slower. Supercomputers are made from commodity components.
What the designer sees on the shelf is what the designer gets. There is
nothing else. Once again, poof there goes flop rate because software
slows performance way down. The oher point:

-->  Commodity parts help flop perfor-      <--
-->  mance but harm simulation performance. <--

Alas, there is silence on these issues.

Paul

Then Paul responds to Doyle's reply,
Flaco:

On Sat, 20 Apr 2002, Doyle Saylor wrote:

Greetings Everyone,
Paul goes on about flops per second in regard to the limitations of
speed and other questions in a super computer.  Which shot my thoughts
toward Steve on this list who was working on quantum computing (QC) for a
short time.  The basic idea being that a single operation in a QC could
produce an answer to some problem.  I'm wondering if that isn't a work
around or patch to the limits of super computing.  But I suppose doing a
simulation isn't about a single operation.  The quantum computer would be
useful for cracking codes, but not so good at the simulation modeling?!?

Paul
QC is beyond my small arena of knowledge. The stuff I'm talking about is
exotic but not THAT exotic. Is there such a thing as ordinary exotic,
maybe plain exotic?

Doyle
I tend to suspect that the arguments about what works in super computing
usually are not subject to popular analysis.  The debates are not carried

It could easily be done in the style of Scientific American. It's just
that no one has written anything. The points I was making are hardly
anything secret and not the subject of debate, though they should be. Hold
that. There was a debate once at a casual level. It was a breakfast last
year with a data base researcher and all around great guy, Jim Gray (James
N. Gray) who is not supposed to know math algorithms but knows a lot.
Mostly the technical points on flops are the result of personal musings
and experience. The points about word length require more technical
investigation before advancing beyond casual breakfast time conversation.

Doyle
out in public because the research is for secret weapon work.  So we don't

Paul
Like I say, it's not that exotic. The stuff on flops and supercomputers
has to do with algebra which may be talked about by people with clearances
but everything I know about in algebra is open published knowledge. Been
that way for years. There are of course technologies that are secret
either in private industry or in military applications. But nothing on
this algebra. But then I repeat myself, repeat myself.

The big point to make is that supercomputers are not the whole story when
it comes to simulation. There are issues that get little attention from
anywhere that limit supercomputer effectiveness. These are algorithmic
issues and are not limitations due to the devices getting to be too small.
Even with a QC that could do algebra if the algorithms are inappropriate
there will not be a correct answer.

Doyle
really know why some technology gets abandoned as the Times article points
out about vector super computers.  The U.S. abandoned super fast torpedoes
and submarines in the sixties.  The Russians continued their research, and
seem to have come up with designs that yield very fast under water
propulsion.   But since this is military information we aren't likely to
know the details for decades if ever.  As to the Japanese, they must have
custom designed their machine which as Paul suggests in his comments in the
U.S. super computers are generally made out of off the shelf computers
components.

Paul
As usual my lack of clarity muddies the picture. The NEC machine is just a
slick high performance wonder, not an exotic architecture. More dependent
on vectorization than US machines as I understand what Dongarra said.
There are custom machines for example trying to compute the mass of the
proton (if I have this correctly stated.)

Doyle
There then is question about secrecy.  In American business the NDA (non
disclosure agreement) is used to conceal secrets that need to be shared
between businesses seeking to market to each other.  Some people, Michael
Perelman at Chico State University in California thinks Intellectual
Property rights gum up technical development.  Military secrets affect such
rather curious things like GPS (global positioning information from
satellites).  The argument being too accurate GPS is national security.
Together business secrets and military secrets potentially become the
determining factor in what happens.  For example, Michael Perelman has
written;

Paul
Some people say that corporate secrets are more carefully locked away than
military secrets. Like the formula for Coca Cola?


Doyle
Michael quoting first Fred Warshofsky's estimate of the cost of merely
searching for possible conflicts in patents;
"A [software] system with 100,000 components, for example, can use hundreds
of previously patented techniques.  Because each patent search costs about a
thousand dollars, searching for all the possible patent potholes in the
program could easily run well over $1 million, and that far exceeds the cost
of writing the program.3"

Paul
A real software system with graphics, realistic I/O compares to the design
of a city and the interactions of all the human components. The algorithms
affecting algebra and simulation are tiny.

Doyle
Writing in the early days of biotechnology, a Business Week article cited an
industry attorney who observed, "there are more lawsuits than products."  At
the time, Genentech sold only two products; it was already engaged in seven
suits.4"

The reason that Japan may continue to advance their technology may also
be the reason that the Chinese do also.  Where Intellectual Property is
gradually tying up U.S. technical knowledge into unshareable bits and
pieces, the Japanese and Chinese actually share information more readily
than the U.S. does internally.  The U.S. complains bitterly about commodity
piracy in China, but that piracy is a sign that the system there is porous
with regard to 'secrets'.  At least this is Michael's theory and has some
interesting flavor to it as far as I am concerned.
thanks,
Doyle

Paul
Interesting stuff to talk about. Way outside my domain.

Gordo